Introduction
Unfortunately, due to uploading my files to the comment section of the first homework assignment, my songs did not appear in the class corpus. I attempted to run my tracks through Essentia using Python, but encountered installation errors. Therefore, for this assignment, I will focus on comparing the class corpus with the AI Song Contest (AISC) dataset.
My research will focus on identifying the outliers in both datasets, particularly in the feature of danceability, while also considering potential factors such as tempo and instrumentalness. I will begin by analyzing these outliers based on the perceived values of these features, without listening to the tracks. After identifying the outliers, I will then listen to the tracks and compare my subjective interpretation of these features to the results from Essentia. This process will allow me to explore how well the feature extraction methods align with my own musical perceptions and whether the outliers identified by Essentia fit within a broader musical context.
Visualizing the Class Corpus
This scatter plot visualizes the Class Corpus dataset, focusing on the relationship between tempo (x-axis) and danceability (y-axis), with instrumentalness represented by color. Tracks with higher tempos tend to have higher danceability, while those with slower tempos and lower instrumentalness are less danceable.
Visualizing the AI Song contest
This scatter plot visualizes the AI Song Contest 2024 dataset, displaying the relationship between tempo (x-axis) and danceability (y-axis), with instrumentalness represented by color.
Comparing Song contest vs Class
Identified Outliers in Both Datasets
Class Corpus (compmus2025):
Least danceable track: erik-l-1
Tempo: 30 BPM
Instrumentalness: 0.1309513
Most danceable track: roemer-i-1
Tempo: 176 BPM
Instrumentalness: 0.9554076
AI Song Contest Dataset (aisc2024):
Least danceable track: Vonpsyche
Tempo: 63 BPM
Instrumentalness: 0.3695597
Most danceable track: Almost Human
Tempo: 155 BPM
Instrumentalness: 0.9040143
Conclusion & Discussion
The relationship between tempo, instrumentalness, and danceability in these tracks reveals some interesting patterns. The least danceable track from the class corpus, “erik-l-1,” stands out with its slow tempo (30 BPM) and relatively low instrumentalness (0.1309513). While the track may lack the vocal presence typically associated with more energetic dance tracks, its low instrumentalness and slow tempo make it feel more like a classical piece, which could be associated with more reserved forms of dancing, like a waltz. While not traditionally considered “danceable,” its rhythm could still engage listeners in a more contemplative, slower-moving way.
“Roemer-i-1,” the most danceable track in the class corpus, shows how the combination of a fast tempo (176 BPM) and high instrumentalness (0.9554076) can elevate a track’s energy. Its beat, which I would categorize as IDM or trip-hop, contributes to its overall danceability. The high tempo paired with intricate percussion and experimental elements makes this track ideal for a more eclectic and energetic dance environment.
In the AI Song Contest dataset, “Vonpsyche” stands out with its low danceability, which I think can be explained by its moderate tempo (63 BPM) and moderate instrumentalness (0.3695597). The track, mostly made up of spoken word over an ambient beat, prioritizes atmosphere over rhythm, making it harder to categorize as a dance track. Its focus on vocal delivery and mood over a driving beat directly impacts its lower danceability score.
Finally, “Almost Human” is the most danceable track in the AI Song Contest dataset. With its tempo of 155 BPM and high instrumentalness (0.9040143), this track showcases the power of electronic dance music. Its four-on-the-floor beat and deep, rhythmic structure make it highly danceable, with its instrumental complexity adding a layer of richness that keeps the energy flowing throughout.